Meta-path Based Heterogeneous Graph Embedding for Music Recommendation

Abstract

The prosperous online music streaming industry makes personalized music recommendation a topic worthy of extensive study. Traditional music recommendation techniques which are based on conventional collaborative filtering or acoustic content features usually sufffer from data sparsity or time-consuming computation problems, respectively. In fact, online music services not only generate listening history for each user but also accumulate a large amount of heterogeneous data including performers, tags, ownerships and so on. Capturing underlying user preference from the heterogeneous data to enhance music recommendation is transparently promising, because on one hand these data can mitigate the sparsity of listening history while incorporating them into recommendation model is computationally affordable. To this end, in this paper we propose a novel music recommendation approach. It first models the music system as a heterogeneous music graph. Then, to make full use of the heterogeneous data, carefully designed meta-paths are used to dig up the information lying in the graph. Finally, we learn user preferences through a combination of Bayesian Personalized Ranking model and heterogeneous embedding representation learning. Extensive experimental analysis on real-world public dataset validates that the proposed approach outperforms the baselines, especially on cold start users.